Generated normal/reflection PROJECTED on actual WebGL1/2

2026-11-04. Run: scratch/runs/20261104-legacy-generated-projected-webgl/. Source revision and all 48 runtime pins are in source-seal.json. This tests the generated normal/reflection repair e6a787467; it does not rerun the accepted camera-position PROJECTED suite.

Result and defect boundary

All 32 positive cells pass on the actual browser graphics contexts: canonical/compat modules × WebGL1/2 × immediate/physical owning Worker × ordinary/indexed × direct/VB sources. This is 736 successful public COM draws and 2,208 exact RGB observations, including 608 generated PROJECTED draws and 1,824 generated pixels. The remaining 128 draws/384 pixels are supplied and XYZRHW bypass controls. Both COUNT3/4, all four diffuse/specular layouts, lighting conversion and explicitly normalized reflection are included.

No positive-path renderer or transport failure occurred. Accordingly this task makes no new production renderer repair claim and changes no production renderer file. It establishes actual WebGL acceptance of the existing repair. The requested conditional stop point was the actual pixel result; there was no first failure to diagnose or justify a production OWN/edit. The browser closed after the controls, with no second launch.

The 26 rendered negative controls reject zeroed normals, identity matrix, supplied selection, constant texture q, lost texture z, explicit per-vertex early division for both modes, and normal selection instead of reflection, on both WebGL versions. Every rejection follows a successful public draw and an actual pixel mismatch at unchanged tolerance zero. Early division yields the predecessor baseline pixels: normal [55,30,201] versus [185,82,19], reflection [107,82,201] versus [107,69,136]. These are intentionally faulty public inputs, not newly discovered production failures.

Fixture and independent checks

test/test-legacy-generated-projected-webgl.js reuses the exported fixtureSource, config and oracle from test/test-legacy-generated-normal-reflection-projected.js, the actual browser lifecycle/HTTP/bootstrap harness from the count-selector runner, and its underlying projected browser fixture. The exported software run supplies a separate two-draw/six-pixel control, not GPU evidence.

The new fixture holds production endpoint batch delivery to the physical owning GL Worker. Across all 16 Worker cells, 368 held records, including 304 PROJECTED records, survive public world/view/texture/selector/flags changes, direct-source/index overwrites and public VB Lock/overwrite/Unlock. Full held bytes compare identically before and after mutation. Captured vertices retain independent camera normal/reflection XYZ1 and geometric reciprocal W. Public target Lock provides the completion/readback fence. Owning Worker statistics show the requested WebGL version and zero software draws, fallback draws, fallbacks and errors. Immediate cells additionally compare actual GL readPixels RGBA with the oracle and check GL error zero.

To avoid 1,024 cross-Worker reads per observation, completed Lock backing is read through a DataView after the public Lock fence. The selected sample still uses the public guest read; the retained full DIB supports inspection. No production transport, synchronization or sampling helper supplies expectations.

The independent Python Fraction audit derives object/world/view geometry, unit eye vectors, reflected vectors, texture U/V/Q, geometric W and texel addresses. It verifies all 2,208 positive pixels, 1,824 distinctions from early division, 368 captured records, all 26 negative mismatches and 26,416 explicit stack cleanup checks among 26,608 public call/release records. Every authored sample retains its texel under ±1/16-pixel displacement in both axes. It uses sum(b*U/w)/sum(b*Q/w) and the corresponding V quotient independently of production matrix or renderer helpers.

Preflight, exact inputs and custody

Before staging, assigned-box exec verified both runtime paths, retained Puppeteer 24.31.0, no Chrome, and healthy fully hydrated runtime. Under the local build lock, exact current modules were transferred directly into runtime build/ paths, with hashes checked before/after. Neither modules nor media were copied into evidence. Existing calc input and dependency paths/hashes are recorded in the seal. Changed small helpers were transferred individually.

Module SHA-256
canonical b70c420e0a29c8664df1c3ed8b8074b590c342ed2a60078172b30e90922800cf
compat 4771c78ba2cc1b84875e1e90a743f7fa624975764607c3d86a516fbcd1cd5d00

CJS/ESM Puppeteer imports, both module validations, actual public COM bootstrap, HTTP served hashes/isolation and all 48 pins passed before Chrome, including a separate successful --preflight. Cache graph verification passes locally (98 page / 13 Worker scripts). The serialized full build and E2E classification checks are recorded in the run. The new browser test is explicitly registered in tools/test-tiers.js so the ordinary unit tier does not launch Chrome.

The sole Chrome 151 process PID 12444 opened at 14:04:09.435Z and closed at 14:04:53.382Z, ahead of the fixed 14:27Z hard close. CDP reports SwiftShader; the context reports Intel/Mesa. This conflicting identity does not qualify physical hardware. Closed inspection found no Chrome, HTTP port 42309 closed, and all 48 runtime pins unchanged. Six unique closed log/result/seal files were recovered and SHA verified against an identical second remote inventory. The large result was gzip-compressed only after exact roundtrip verification.

Assigned bx_vvqgqtk6 is left RUNNING / API idle, healthy and fully hydrated, expiry 2026-11-04T14:41:01.315Z. Custody returned at 14:06Z; the coordinator owns normal stop at 14:37Z. No lifecycle operation, other box, local Chrome, subworker, game, deployment or timer extension occurred.

Recovery uses Git SHA, relevant binary diff and owned new-file patches; there is no source archive or duplicated media/module evidence. PROJECTED remains unadvertised. COUNT2, clipping, nonrigid normals, strided generation, native ProcessVertices output, native Windows, full27 and gameplay remain unqualified. Next is separately scoped native ProcessVertices evidence or a distinct demonstrated supported rendering failure; this single task stops here.